Gemini 4 Neon vs. Claude Opus 5.5 Compared on Coding, Speed, and Price
TL;DRStart with Gemini 4 Neon for routine coding if cost is your main consideration. Google released Neon on November 18. Its standard API rates for input, output, and cached reads are 25% below Claude Opus 5.5’s. Vals AI ran both models through the same terminal agent. Opus passed 65.15% of the tests against Neon’s 61.62%, while Neon finished sooner and cost less per test. For a difficult CLI repair or a migration that blocks other work, Opus is worth trying first.
By Cameron Voss and West Mao
Gemini Neon vs. Opus at a glance
Neon sits between Hydrogen and Carbon in the Gemini 4 lineup. Google positions it for complex tasks with shorter turnaround. Anthropic released Opus 5.5 on September 22 for long-running coding agents and knowledge work. Both have a roughly one-million-token context window. Their supported inputs and tool behavior differ.
| Dimension | Gemini 4 Neon | Claude Opus 5.5 |
|---|---|---|
| API model ID | gemini-4-neon | claude-opus-5-5 on the Claude API |
| Context window | 1,048,576 tokens shared across input and output | 1 million tokens |
| Synchronous output | 65,536 tokens per foreground segment. Supported APIs can continue across segments within the shared context budget. | Up to 128K output tokens on the standard Claude API |
| Inputs | Text, images, PDFs, audio, and video | Text, images, and PDFs on supported endpoints |
| Output | Text, including structured output on supported endpoints | Text, including structured output on supported endpoints |
| Default reasoning | High. Low, medium, and extra high are also available. | Medium. Adaptive thinking is always on, with effort adjustable up to max. |
| Where to try it | Gemini API and Google AI Studio, Gemini app and Antigravity for eligible Pro and Ultra users, Google Cloud subject to account and regional availability | Claude Code, Claude API, and supported AWS, Google Cloud, and Microsoft Foundry deployments |
| Standard API input / output / cache read | $3 / $15 / $0.15 per million tokens | $4 / $20 / $0.20 per million tokens |
Google Cloud and Anthropic publish the model limits and API settings above. Standard API rates are in USD.
Coding benchmarks and what they measure
Neon leads on CursorBench 4.1 in Google DeepMind’s November 18 comparison. Opus has a smaller lead on Terminal-Bench 4.1 and FrontierCode 2.0. DeepMind combined its own runs with external results. Compare the models within each row because the tests use different tasks, tools, and scoring rules.
| Benchmark | Gemini 4 Neon | Claude Opus 5.5 | What the result captures |
|---|---|---|---|
| CursorBench 4.1 | 65.4% | 58.1% | Performance on coding-agent tasks drawn from real Cursor sessions |
| Terminal-Bench 4.1 | 65.2% | 65.8% | Terminal tasks, with scores adjusted for rule violations. DeepMind combined results from different sources. |
| FrontierCode 2.0 | 46.2% | 47.3% | How often generated code changes meet a mergeability standard |
| SWE-sweep, provider harness | 39.0% | 10.1% | End-to-end performance with Antigravity CLI for Neon and Claude Code for Opus |
| SWE-sweep, Mini-SWE harness | 12.8% | 3.6% | The same benchmark with a shared agent setup. The gap narrows from 28.9 to 9.2 percentage points. |
What happened when both models used the same terminal agent?
Vals AI ran both models through Mini-SWE-agent on Terminal-Bench 4.0. Its verifier checks the final artifact or environment and awards credit only when the full test passes. One operations task asks the agent to speed up worker restarts while preserving timely, duplicate-free notifications during a rolling deployment. A patch can look sensible in review and still fail that check.
| Vals AI Terminal-Bench 4.0 | Gemini 4 Neon | Claude Opus 5.5 |
|---|---|---|
| Pass rate | 61.62% | 65.15% |
| Cost per test | $6.42 | $12.04 |
| Duration per test | 36 min 6 sec | About 64 min |
Opus led by 3.53 percentage points. Neon cost about 47% less per test and averaged just over 36 minutes, compared with about 64 minutes for Opus. That makes Neon a sensible first pass for a batch of tasks that are cheap to verify. A migration that blocks a release may justify paying more for the higher pass rate. Keep the Vals AI 4.0 results separate from the 4.1 scores above, which come from a different setup.
What does a working patch cost?
Using Google’s and Anthropic’s published API rates, 100,000 uncached input tokens and 20,000 billable output tokens cost $0.60 on Neon and $0.80 on Opus 5.5. Each example below holds token counts constant and excludes cache creation or storage, tool fees, and retries. Billable output includes reasoning where the provider charges for it.
| Illustrative API request | Neon | Opus 5.5 |
|---|---|---|
| 100K uncached input and 20K billable output | $0.60 | $0.80 |
| 300K uncached input and 30K billable output | $1.35 | $1.80 |
| 90K cache-hit input, 10K uncached input, and 20K billable output | $0.3435 | $0.4580 |
Anthropic charges $0.20 per million tokens for Opus cache reads. Writes cost $5 per million tokens for a five-minute cache and $8 for a one-hour cache. Check how many tokens actually hit the cache on the second and third calls. An agent that reuses a repository prefix can spend much less than one that rebuilds its context on each turn. One accepted Opus patch can also cost less than two Neon attempts, so track total spend per accepted patch.
Which model is faster at its default setting?
Artificial Analysis measures when the first visible answer token arrives and how long a single request takes to finish. Its separate time-per-index-task estimate combines benchmark token use and measured output speed. The table shows the first two measures. For a coding agent, the time until a patch passes CI is a separate measurement to capture in your own test.
| Reported first answer token / mean single-request completion | Gemini 4 Neon | Claude Opus 5.5 |
|---|---|---|
| Low effort | 13.93 s / 32.85 s | 5.47 s / 51.15 s |
| Medium effort | 20.14 s / 74.51 s | 14.78 s / 78.92 s |
| High effort | 40.64 s / 81.29 s | 24.55 s / 101.14 s |
| At documented defaults, Neon high and Opus medium | 40.64 s / 81.29 s | 14.78 s / 78.92 s |
At their documented defaults, the mean completion times in this test were about 2.4 seconds apart. Opus started answering about 26 seconds sooner. Artificial Analysis also found that raising Neon from high to extra high moved its Intelligence Index score from 60 to 61 and raised estimated cost per index task from $3.74 to $6.44. Test that setting on tickets that regularly fail at high effort before using it for every run.
Long context, multimodal work, and agent integration
How well do they use a long context?
In DeepMind’s GraphWalks evaluation, Neon scored 95.2% F1 against Opus 5.5’s 66.8% on 200 graph-traversal problems with inputs from 256K to 1M tokens. You can test a similar skill in your own repository. Ask each model to trace a behavior from its caller through shared configuration to a downstream test. Check every file reference and leave room in the context window for tool results and the final diff.
Which can work with screenshots and recordings?
Both accept screenshots and PDFs. Neon also takes audio and video as input, which helps when a bug only makes sense after watching the sequence of events. With Opus, give it selected frames and a transcript alongside the code. Artificial Analysis’s knowledge-work tests split by task. At Neon’s extra-high setting and Opus’s max setting, Neon scored 79.3% against 69.5% on AutomationBench-AA, while Opus led AA-Briefcase v1.1 by 1,800 to 1,699. For your own trial, choose a deliverable you can check, such as a spreadsheet with formulas you can recalculate.
What changes when you switch a production agent?
Test Neon in Antigravity or in an existing Gemini API integration that uses Google grounding and function calling. If the team already uses Claude Code, Opus 5.5 is easy to trial there. Anthropic’s migration guide says the Opus API rejects forced tool selection. Check custom integrations that require a specific tool call on every request. For private repositories, verify retention and cache settings for the account and endpoint you plan to use.
Which model should your team use?
| If your main job is… | Start with… | Check before rolling it out |
|---|---|---|
| Routine bug fixes, repository exploration, and broad test additions | Gemini 4 Neon | Hidden regression tests, unnecessary file changes, and cost per approved patch |
| A difficult CLI repair or multi-step migration | Claude Opus 5.5 | Full CI pass, backward compatibility, and recovery after a partial failure |
| A fix driven by a screen recording, logs, and screenshots | Gemini 4 Neon | Whether the agent identifies the correct sequence of events and can verify the fix |
| An established Claude Code workflow | Claude Opus 5.5 | Median completion time and reviewer edits at the default effort |
| A customer-facing workflow with a tight response deadline | Trial both at matched budgets | Time to first visible response, time to completion, and tail latency on your traffic |
Pick ten resolved tickets that cover small fixes, multi-file refactors, and difficult terminal or migration work. Run both models from the same pre-fix commit with the same tools, time limits, and acceptance tests. Run a second round in Antigravity and Claude Code if those are the tools your team would actually use. Count accepted patches, then record spend, time to passing CI, and reviewer corrections. This gives you a local comparison to put alongside the public benchmark results.